BRADS: A Multipurpose Audio Dataset For Bangla Regional Word Detection

Research Hypothesis
This dataset tests the hypothesis that Bangla ASR performance is affected by regional dialects and pronunciation variations. Despite Bangla’s widespread use, speech recognition models struggle with dialect diversity. This dataset enables the development of more accurate and inclusive ASR models.

DATA SUMMARY
The dataset contains 298 Bangla words (233 regional, 65 chaste Bangla) recorded by 85 native speakers from eight divisions, resulting in 2,439 high-quality audio samples.

Data formats include .wav (audio files) and .xlsx (text data). Words were recorded using recommended apps, verified manually, and include background noise for real-world ASR training.

NOTABLE FINDINGS
Pronunciation varies significantly across regions. For example, the word "আমি" (Ami) [I] differs:

Chittagong: "আই" (Ayi)
Barisal: "মুই" (Mui)
Rajshahi: "আমাক" (Amak)
Rangpur: "হামি" (Hami)

Chittagong has the most dialectal variance, while Rangpur is closest to chaste Bangla. Most contributions came from ages 23-27, indicating generational trends.

HOW TO USE THE DATA
The dataset can be used for ASR training in CNN, RNN, and Transformer models. It also improves NLP applications, chatbots, and speech-to-text systems. Linguists can study phonetic variations, and it enhances Bangla-English machine translation.

VALUE OF THE DATA
This dataset fills a gap in Bangla ASR research, supporting inclusive AI development and linguistic diversity preservation. It serves as a benchmark for Bangla speech-based AI, making voice technology more accessible